You are on page 1of 8

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015

RESEARCH ARTICLE

OPEN ACCESS

Multifractal-Based Featuresfor Medical ImagesClassification


Saad Al-Momen1 , Loay E. George2, Raid K. Naji3
Computer Science Department, College of Science,
Baghdad University, IRAQ
----------------------------------------************************----------------------------------

Abstract:
This paper presents a method to classify colored textural images of skin tissues. Since medical images have
highly heterogeneity, the development of reliable skin-cancer detection process is difficult, and a mono fractal
dimension is not sufficient to classify images of this nature. A multifractal-based feature vectors are suggested here
as an alternative and more effective tool. At the same time multiple color channels are used to get more descriptive
features.
Two multifractal based set of features are suggested here. The first set measures the local roughness property, while
the second set measure the local contrast property.A combination of all the extracted features from the three color
models gives a highest classification accuracy with 99.4048% for training and 95.8333% for testing.

Keywords:-Texture Classification, Texture Analysis, Fractal, Multifractal, Wavelet Features


----------------------------------------************************---------------------------------I.

INTRODUCTION

A TUMOR IS RECOGNIZED AS THE EXISTENCE OF


AN ABNORMAL MASS OF TISSUE WITH A CAPACITY FOR
PROGRESSIVE GROWTH. IT IS A TERM USED TO DESCRIBE
BODY CELL CHAOTIC GROWTH AND DIVISIONS
OCCURRING IN AN UNCONTROLLABLE FASHION,
USUALLY DUE TO CELL DNA CHANGE OR DAMAGE [1].
TUMORS CAN BE CLASSIFIED INTO TWO MAIN CLASSES
OF BENIGN OR MALIGNANT [2]:

Benign (not cancer): Benign tumors are rarely lifethreatening, and theydo not spread to other parts of
the body. They often can be removed andusually do
not grow back.
Malignant (cancer): Malignant tumors can harm
nearby tissues andspread to other parts of the body.

According to the World Health Organization


(WHO), malignant tumor or cancer is the leading causes
of morbidity and mortality worldwide, with
approximately 14 million new cases and 8.2 million
cancer related deaths in 2012. The number of new cases
is expected to rise by about 70% over the next 2 decades
[3]. There are as many different types of tumors as there
are different types of human cells, just over 200 types,
with some being very common, while others are
extremely rare [2]. Nearly all tumors are named after the
organ or type of cell that they originate from.

ISSN :2394-2231

Cancer arises from one single cell that takes a


multistage process to transform from a normal cell into a
tumor cell, typically a progression from a pre-cancerous
lesion to malignant tumors. These changes are the result
of the interaction between a person's genetic factors with
the physical, chemical, and biological carcinogens.
Cancer mortality can be reduced if cases are
detected and treated early. The appearance of body
organs in addition to other medical tests, such as biopsy
specimens, body fluid analysis and measurement of
body functions, can be significant for physicians to
reach to a decision on the medical situation of the patient.
Physicians can have an initial idea on the normality
orabnormality of the observed organ from its general
appearance via an image captured. Image texture is one
of the important cues that could give physicians such
indication, which would trigger certain treatment
procedures, if texture is found abnormal, depending on
the nature of the disease.
To avoid diagnosing disease at an advanced stage
when it has already progressed and hence patients
prognosis becomes poor, early detection of disease can
be improved by using effective clinical diagnosis
systems.

II. SKIN CANCER


Skin cancers are named for the type of cells
where the cancer starts. It is also known as skin

http://www.ijctjournal.org

Page 6

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015

ISSN :2394-2231

Class 3:
BCC

Class 2:
Benign

Class 1:
Normal

cancer or the grade of the cancer which makes


image analysis for tissue and cell images is
complicated and challenging. Therefore, one has to
generalize the analysis with applying more
advanced mathematical techniques.
This paper analysis the skin tissue medical
images from the multifractal point of view. Many
researchers conclude that multifractal technique is
an effective and robust tool for image classification
[6], [7], [8], [9], [10] and [11]. Therefore,
multifractal technique has been widely applied in
biomedical image processing. Multifractal describes
the fractal properties of an image using an intensitybased measure within the neighborhood of each
pixel. Although there is a link between roughness
and fractal dimension, the roughness is not
sufficient to describe a textured surface, because
other characteristics have to be involved, such as
arrangements and spatial distribution of grey levels.
Multifractal theory can avoid the drawback of the
single fractal dimension.
In multifractal, instead of one measure, ?,
describing the phenomenon in all scales using
fractal approach, the set of measures, ???i, arise,
describing statistically the same phenomenon in
different scales.

Class 4:
SCC

neoplasia. Skin cancer is a change in some of the


cells of skin such that they grow abnormally to
form a malignant tumor. These abnormal cells can
invade through the skin into adjacent structures or
travel throughout body and become implanted in
other organs and continue to grow; a process called
metastasis [4].
There are three common types of skin cancer,
basal cell carcinoma (BCC), squamous cell
carcinoma (SCC) and melanoma. There are other
types, but they considered more rare forms of skin
cancer. BCC and SCC are the most common forms
of skin cancer; and they are together referred to as
non-melanoma skin cancer. While, Melanoma is
generally the most serious form of skin cancer
because it tends to spread (metastasize) throughout
the body quickly [5].
Figure 2 shows sample examples of the skin
tissue classes that used in this study.These images
were taken from stained sections (glass slides) of
skin cancer biopsy cases from different Iraqi
hospital pathology departments. Taking images
from these sections was conducted through using
microscope-attached digital camera; it was
performed with the help of an experienced
pathologist. The pathologist was required to pinpoint the areas of interest in histological [4].It is
clear that the tissue structure is not identical in all
image samples and varies from one patient to
another, since disease might alter the tissue
structure unequally. This tissue heterogeneity
places tissue patterns commonly in the category of
stochastic or possibly fractal textures.
Generally speaking, the image textures of humanorgan tissues are complicated as anything in nature,
and digital microscopic images of these tissues
show highly irregular texture patterns with the
variation of the image resolution. It is possible that
different types of texture may have the same fractal
dimension; so, it is difficult to realize the
pathological changes located in different organs
only by methods based on single fractal, especially
or the textures that possess similar fractal
dimension.
Also, the fractal dimension of such type of
texture images is not constant over all scales, but
rather over small ranges of scales. Furthermore, the
cell's texture patterns vary with the type of the

Figure1.Sample examples of the skin tissue classes

III. PROPOSED SCHEME


To classify the skin cancer tissues, two
multifractal based set of features are suggested. The first

http://www.ijctjournal.org

Page 7

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015


set measures the local roughness property, while the
second set measure the local contrast property. These set
of features are extracted from multiple color channels in
order to get more informative feature vectors and to
cover the heterogeneity of the studied textures.Three
color channels
, ,
were taken for each image in
addition to their corresponding brightness (gray). Three
color systems were studied here: RGB, Lab,
Lab and HSV
[12] and [13]. According to that, the four channels will
be taken in this study are (R, G, B, g) , (L
L, a, b, g), and
(H, S, V, g) corresponding to the RGB, Lab,
Lab and HSV
system, respectively.
Figure (2) shows that the input image
ima is passed
through the following main steps:
1- Decomposing the original image into 3 color
channels and 1 gray channel depending on the
chosen color system.
2- For each one of the four channels, two feature
vectors are calculated:
a. The local roughness feature vector F1.
b. The local contrast feature vector F2.

d T ,F

(1)

Where, T is the template value of kth feature that


belong to ith class; F is the value of kth feature extracted
from jth sample; is the standard devotion over the
sample set.
As mentioned above, the matchin
matching process uses three
templates per class, in order to maximize the probability
of true match classification and minimize the
misclassification. The efficiency of classification is
calculated for each distance using the following equation
[16]:
%

$%&' ($.$* +&,-'.+ /$.$* , +0'&++ 1 .2 +&,-'.+


$%&' ($.$* +&,-'.+

4 100%

3- The two feature vectors F1 and F2 are used


individually or to gather to establish the
classification task.
During the training phase, three templates are
constructed for each class to cope over the variability
variabi
of
the images in each class. The conducted experiments
showed that using one or two templates may not always
enough for efficient classification. In this research paper,
three initial templates have been chosen as: (i) IC ,
which is the mean featuree vector of all the feature
vectors extracted from the training samples belong to the
class, (ii) IC , the farthest feature vector to IC , and (iii)
IC , the farthest feature vector to both IC &IC . Then
the K-means algorithm is used to improve
rove the values of
these initial templates [14], and [15].
Commonly, Euclidean distance measure is used to
match the similarity. But, one weakness of the basic
Euclidean distance function is that if one of the input
features has a relatively large range, then
t
it can
overpower the other features. Since the problem here is
the used features are not isotropic; that is, every feature
may not have similar behaviors. So, the normalized
Euclidean distance has been used to evaluate the
similarity degree between the extracted feature vector of
the tested sample, and the templates representing certain
class [14] and [15]:

IV. FEATURES EXTRACTION


Two multifractal based set of features are suggested here.
The first set measures the local roughness property,
while the second set measure the local contrast property.
A. Local Roughness Features

Roughness is a component of surface texture. It is


measured by the deviations in the direction of the
normal vector of a real surface from its ideal form. The
surface is considers as rough surface when the
deviations are
re large; otherwise it is smooth.
The human-organ
organ tissues have high heterogeneity
nature, so the roughness is different from part to part.
Local roughness features are suggested here to cover this
issue. This set of features depends on the local variations

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015


in the pixels' values relative to their local position in the
image. The original image of size m4n is divided into
blocks of size b4b, where b is an odd number. The
differences between the pixel's value in the block's
center and the pixels surrounding it considered as local
roughness measure. After the local pixels differences are
computed for each central pixel, two matricesLRmin()
8
;
and LRmax() of size 7 9 : 4 79 :, are constructed to hold
the minimum and maximum of these differences
respectively.Then the fractal dimension for each of
LRmin() and LRmax() are computed using the BCABH
method[17]to produce two values FD(LRmin) and
FD(LRmax); Figure (3)illustrates these steps.
The above procedure is repeated to all the color's
channel , ,
and g to get a feature vector (F1) that
consists of eight values.
<

{<>?@ ABCDE , <>?@ ABCFG , <>?3 ABCDE , <>?3 ABCFG ,


<>?H ABCDE , <>?H ABCFG , <>I ABCDE , <>I ABCFG }

with R8ST and R8U; representing the highest and lowest


luminance.
A multifractal technique is used here, where a window of size
w is centered at a pixel p and the contrast measure with
respect to w can be characterized as follows:
VW X
>X YZ
X

(5)
2D + 1,

0, 1, 2, , ^

(6)
whered is the total number of windows used to compute _W .
_W log X + log >
log VW
(7)
In other word,
c _W G + d
(8)
where c log VW , G log X and d log > .
It is clear from eq. (8) that _W is the slope of the line, and it
can be estimated using a linear regression
; Tf T f
_W
3
; T

T T

(9)
Figure (4) illustrates the main steps to calculate the local
contrast features, where:

1- The original image is divided into block of size


b4b, where b is an odd number.
2- For each block, _W is calculated for the central
pixel according to a predefined number of
windows d. And the result is set in a matrix
Con().
3- The fractal dimension of Con() is calculated
using the BCABH method [17].
Figure 3. Fractal Dimension of Local Roughness

V. LOCAL CONTRAST FEATURES


Contrast is
the
difference
in luminance or color that
makes
an
objectdistinguishable. In visual perception of the real
world, contrast is determined by the difference in
the color and brightness of the object and other objects
within the same field of view.
There are many possible definitions of contrast.
Some include color; others do not. Michelson contrast
[18] is considered here because it is usually used for
patterns where both bright and dark features are
equivalent and take up similar fractions of the area,
which is the case of the studied tissues. The Michelson
contrast is defined as

The above procedure is repeated to all the color's


and g to get a feature vector (F2) that
channel , ,
consists of four values.
<

{<>?@ dgE , <>?3 dgE , <>?H dgE , <>I dgE }


(10)

KLMN KLOP
KLMN QKLOP
(4)

ISSN :2394-2231

VI. EXPERIMENTAL RESULTS

http://www.ijctjournal.org

Page 9

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015


In order to demonstrate the efficiency of the
proposed classification system for the medical images, a
number of experiments have been performed. 24 skin
tissues are predefined by a pathologist into 4 main
classes: 6 are Normal, 6 are Benign, 6 are Basal Cell
Carcinoma (BCC) and 6 are Squamous Cell Carcinoma
(SCC). Each tissue defines a separate subclass from the
corresponding main class, and it is represented by 30
randomly selected colored images of size 128 x 128
pixels.7 randomly chosen samples from each class were
used for training, while the rest 23 samples were used
for testing the classifier. The conducted tests have been
directed toward finding to which class the query image
belongs.
The classification was in two stages. The first stage does
the classification on the basis of 24 sub-classes. And in
the second stage each sub-class is assigned to its
corresponding main class.
Tables(1), (2) and (3) show the percentage of
correctly classified samples of all the tested samples
under the use of the local roughness feature vector (F1),
depending on the RGB, Lab and HSV color model,
respectively.
With the RGB model the highest percentage of
correctly classified sample achieved with block of
size b=17, and the values are 100% for training and
95.6522% for testing.
With the Lab model the highest percentage of
correctly classified sample achieved with block of
size b=9, and the values are 97.619% for training
and 92.3913% for testing.
With the HSV model the highest percentage of
correctly classified sample achieved with block of
size b=9, and the values are 97.619% for training
and 88.7681% for testing. At the same time there is
a higher result with b=7, when excluding the gray
channel, the accuracy values will be 98.2143% for
training and 91.4855% for testing.
Tables (4), (5) and (6) show the percentage of
correctly classified samples of all the tested samples
under the use of the local contrast feature vector (F2),
depending on the RGB, Lab and HSV color model,
respectively.
With the RGB model the highest percentage of
correctly classified sample achieved with number of
windows d=1, and the values are 85.119% for
training and 65.5797% for testing.
With the Lab model the highest percentage of
correctly classified sample achieved with number of
windowsd=1, and the values are 88.0952% for
training and 62.8623% for testing. At the same time

ISSN :2394-2231

there is a higher result with d=1, when excluding


the gray channel, the accuracy values will be
99.4048% for training and 92.2101% for testing.
With the HSV model the highest percentage of
correctly classified sample achieved with number of
windowsd=3, and the values are 95.8333% for
training and 86.0507% for testing.

Tables (7), (8) and (9) show the percentage of


correctly classified samples of all the tested samples
under the use of the local roughness and local contrast
feature vectors together {F1, F2}, depending on the
RGB, Lab and HSV color model, respectively. The
block size for the local roughness is fixed to b=9.
With the RGB model the highest percentage of
correctly classified sample achieved with number of
windows d=1, and the values are 98.8095% for
training and 88.587% for testing. At the same time
there is a higher result with d=3, when excluding
the gray channel, the accuracy values will be
99.8048% for training and 89.1304% for testing.
With the Lab model the highest percentage of
correctly classified sample achieved with number of
windows d=3, and the values are 98.2143% for
training and 90.942% for testing.At the same time
there is a higher result with d=1, when excluding
the gray channel, the accuracy values will be 100%
for training and 92.2101% for testing.
With the HSV model the highest percentage of
correctly classified sample achieved with number of
windows d=3, and the values are 95.8333% for
training and 86.0507% for testing.
Tables (10), (11), (12), and (13) show the classification
accuracy depending on the possible combination of the
color models. A combination of all the extracted features
from the three color models gives a highest classification
accuracy with 99.4048% for training and 95.8333% for
testing.

VII. CONCLUTIONS
According to the tests results presented in this
paper, the following conclusions have been derived:
The Local Roughness Feature Vector work better
with the RGB color model. While The Local
Contrast Feature Vector work better with the Lab
color model.
Excluding the gray channel from the RGB model
enhance the performance of the Local Roughness
Feature Vector classification. And excluding it

http://www.ijctjournal.org

Page 10

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015

from the Lab model enhance the performance of the


Local Contrast Feature Vector classification.
A combination of the Local Roughness Feature
Vector and the Local Contrast Feature Vector
works better with the Lab model.
Although the Local Contrast Feature Vector does
not give a high accuracy classification but it
supports the Local Roughness Feature Vector to
raise the classification accuracy.
A Combination of all the extracted features from
the three color models gives highest classification
accuracy as it is clear from Table (13).

VIII. REFERENCES
[1] Louis, C. J.; "Tumors: Basic principles and clinical
aspects", Longman Group Limited, 1978.
[2] Sankar , D.; "Fractal Based Techniques for
Classification of Mammograms and Identification of
Microcalifications", PhD thesis, Cochin University of
Science and Technology, India, 2011.
[3] Stewart, B. W.; and Wild, C. P.; "World Cancer
Report 2014", World Health Organization (WHO),
2014.
[4] Abdul-Wadood, D. N.; George, L. E.; and Rashhed,
N. A.; "Diagnosis of Skin Cancer Using Image
Texture Analysis", International Journal of Scientific
& Engineering Research, vol. 5, issue 6, June 2014.
[5] Siegel, R., Ma, J., Zou, a., and Jemal, A., "Cancer
Statistics", A Cancer Journal for Clinicians, Vol. 64,
No. 1, January/ February, 2014.

Computing Science and and Mathematics, Ume


University, Sweden, 2007.
[10] Lopes, R.; Betrouni, N.; "Fractal and Multifractal
Analysis: A Review", Medical Image Analysis, Vol.
13, Pp. 634-649, 2009.
[11] Wang, Z., and Du, P.; "Multifractal Analysis and
Modeling of Chaotic Channels", Journal of Software,
vol. 7, no. 3, March 2012.
[12] Gonzalez, R. C.; Woods, R. E.; and Eddins, S. L.;
"Digital Image Processing Using MATLAB", Second
Edition, Gatesmark Publishing, 2009.
[13] Koschan, A., and Abidi, M.; "Digital color image
processing", John Wiley, 2008.
[14] Al-Momen, S.; George, L.; Naji, R.; "The use of
Gradient Based Features for Woven Fabric Images
Classification", British Journal of Mathematics and
Computer Science, Vol. 6, Issue 1, Pp. 68-78, 2015.
[15] Al-Momen, S.; George, L.; Naji, R.; "Texture
classification using spline, wavelet decompositionand
fractal dimension", Applied and Computational
Mathematics, 4(1), Pp. 5-10, 2015.
[16] Bhiwani, R. J.; Khan, M. A.; Agrawal, S. M.;
"Texture BasedPattern Classification", International
Journal of Computer Applications, Vol. 1, No. 1, Pp.
54-56, 2010.
[17] Long, M.; and Peng, F.; "A Box-Counting Method
with Adaptable Box Height for Measuring the Fractal
Feature of Images"; Radio Engineering, Vol. 22, No.
1, Pp. 208-213, April 2013.
[18] Michelson, A.; "Studies in Optics", University of
Chicago Press, 1927.

[6] Pentland, A.; "Fractal-based description of natural


scenes", IEEE Transactions on Pattern Analysis and
Machine Intelligence, 6 (6), pp. 661674, 1984.
[7] Anh ,V. V.; Maedal , J.; Tieng, M.; and Tsui, T.;
"Multifractal Texture Analysis and Classification",
IEEE1999, Pp.445-449, 1999.
[8] Reljin, I. S., and Reljin, B. D.; "Fractal geometry and
multifractals in analyzing and processing medical
data and images", Archive of Oncology, 10(4), pp.
283-93, 2002.
[9] Nilsson, E.; "Multifractal-based Image Analysis with
applications in Medical Imaging", Masters Thesis in

ISSN :2394-2231

http://www.ijctjournal.org

Page 11

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015


Table 1: The percentage of correctly classified samples under the local roughness feature vector (F1) using the RGB color model
b

R-Channel

G-Channel

B-Channel

RGB-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

75

53.2609

76.1905

55.7971

70.8333

52.3551

97.619

84.6014

78.5714

55.9783

97.619

86.413

67.8571

53.0797

72.0238

56.3406

73.2143

54.1667

96.4286

82.7899

73.2143

54.529

97.619

83.6957

72.619

55.6159

73.8095

55.4348

73.8095

51.087

95.2381

84.4203

73.8095

54.529

97.0238

84.6014

75

53.2609

76.1905

55.7971

70.8333

52.3551

97.619

84.6014

78.5714

55.9783

97.619

86.413

11

67.8571

50.1812

75

54.7101

70.2381

53.0797

97.0238

84.6014

76.7857

51.6304

97.619

83.3333

13

63.6905

49.8188

72.619

54.8913

64.2857

53.442

97.619

84.2391

73.8095

51.2681

98.2143

85.1449

15

66.6667

50.7246

78.5714

58.8768

78.5714

61.0507

98.8095

93.4783

64.2857

49.2754

99.4048

95.2899

17

64.881

44.7464

76.7857

57.2464

81.5476

60.3261

98.2143

94.5652

69.0476

50

100

95.6522

19

66.0714

45.1087

76.1905

56.5217

77.381

55.6159

98.2143

94.0217

72.0238

53.2609

98.8095

94.9275

Table 2: The percentage of correctly classified samples under the local roughness feature vector (F1) using the Lab color model
b

L-Channel

a-Channel

b-Channel

Lab-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

75

65.7609

77.381

63.0435

75

58.1522

97.619

87.6812

66.6667

57.4275

98.2143

90.5797

77.9762

60.8696

80.3571

66.3043

75.5952

59.7826

95.2381

88.7681

73.2143

54.529

95.8333

90.942

77.9762

63.0435

76.7857

63.2246

77.381

60.6884

95.8333

89.3116

73.8095

54.529

96.4286

89.4928

76.7857

62.3188

78.5714

64.4928

73.8095

63.2246

97.0238

89.6739

78.5714

55.9783

97.619

92.3913

11

73.8095

49.2754

82.7381

58.8768

69.4629

54.7101

97.619

88.2246

69.6429

48.7319

97.619

87.6812

13

72.0238

53.8043

77.381

55.2536

75

50.9058

98.2143

85.1449

67.2619

50

97.0238

87.1377

15

69.0476

53.8043

75.5952

54.1667

75

54.7101

98.8095

85.8696

64.2857

49.2754

98.2143

87.6812

17

35.119

27.3551

57.1429

45.1087

69.0476

52.5362

82.1429

72.2826

69.0476

50

97.619

86.413

19

36.9048

27.7174

57.1429

47.4638

69.6429

51.4493

80.9524

71.3768

72.0238

53.2609

99.4048

88.0435

Table 3: The percentage of correctly classified samples under the local roughness feature vector (F1) using the HSV color model
b

H-Channel

S-Channel

V-Channel

HSV-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

47.619

42.2101

61.9048

51.2681

71.4286

51.2681

85.119

72.4638

72.0238

61.9565

95.8333

87.3188

82.7381

61.5942

77.9762

56.8841

68.4524

48.3696

98.2143

91.1232

73.2143

54.529

97.619

85.3261

79.1667

64.4928

73.8095

55.7971

67.2619

52.5362

98.2143

91.4855

73.8095

54.529

96.4286

76.8116

83.3333

64.6739

75.5952

58.5145

69.0476

49.0942

97.619

91.2101

78.5714

55.9783

97.619

88.7681

11

75

61.2319

75

58.7868

73.2143

49.4565

97.619

89.1304

76.7857

51.6304

96.4286

86.5942

13

76.1905

63.0435

75

56.3406

72.4286

50.1812

98.2143

88.4058

73.8095

51.2681

97.0238

86.0507

15

76.1905

59.6014

75

51.6304

72.619

51.2681

95.8333

86.7754

71.4286

52.7174

97.619

85.3261

17

35.119

27.3551

57.1429

45.1087

69.0476

52.5362

82.1429

72.2826

69.0476

50

96.4286

76.8116

19

36.9048

27.7174

57.1429

47.4638

69.6429

51.4493

80.9523

71.3768

72.0238

53.2609

97.0238

78.442

Table 4: The percentage of correctly classified samples under the local contrast feature vector (F2) using the RGB color model
d

R-Channel

G-Channel

B-Channel

RGB-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

55.9524

40.3986

55.3571

48.3696

48.2143

32.4275

81.5476

64.1304

64.2857

40.942

85.119

65.5797

45.8333

41.8478

51.7857

43.2971

42.8571

29.1667

75

55.7971

50

36.7754

82.7381

57.0652

59.5238

40.0362

59.5238

40.0362

37.5

28.8043

67.2619

44.7464

47.619

36.9565

69.6429

47.8261

47.619

36.2319

50

37.6812

35.7143

27.3551

57.1429

46.9203

47.619

36.413

73.8095

46.1957

49.4048

39.1304

38.0952

32.971

37.5

26.6304

65.4762

40.5797

40.4762

36.7754

70.8333

43.6594

Table.5: The percentage of correctly classified samples under the local contrast feature vector (F2) using the Lab color model
d

L-Channel

a-Channel

b-Channel

Lab-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

51.1905

37.1377

56.5476

41.1232

57.1429

41.1232

99.4048

92.2101

64.2857

40.942

88.0952

62.8623

41.0714

37.5

33.1522

50

54.7619

38.587

73.2143

52.7174

50

36.7754

82.1429

59.9638

45.2381

36.413

40.4762

33.6957

53.5714

37.8623

67.8571

48.5507

47.619

36.9565

75.5952

51.6304

42.2619

28.6232

39.881

29.529

47.619

35.8696

64.2857

44.2029

47.619

36.413

71.4286

46.558

32.7381

26.6304

39.2858

25.3623

46.4286

38.4058

57.7381

40.2174

40.4762

36.7754

70.2381

47.4638

ISSN :2394-2231

http://www.ijctjournal.org

Page 12

International Journal of Computer Techniques - Volume 2 Issue3, May - June 2015


Table 6: The percentage of correctly classified samples under the local contrast feature vector (F2) using the HSV color model
d

H-Channel

S-Channel

V-Channel

HSV-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

56.5476

52.5725

39.881

34.4203

55.9524

40.5797

85.119

66.1232

64.2857

40.942

96.4286

85.6884

52.9762

36.5942

34.5238

29.1667

41.6667

41.3043

81.5476

61.413

50

36.7754

80.3571

65.5797

48.8095

34.7826

35.7143

26.6304

53.5714

38.9493

79.7619

58.3333

47.619

36.9565

84.5238

60.3261

48.8095

35.3261

34.5238

25.5435

48.8095

38.7681

71.4286

46.3768

47.619

36.413

75

52.1739

38.6905

30.4348

35.119

26.2681

47.0238

37.6812

63.6905

41.6667

40.4762

36.7754

75

43.6594

Table 7: The percentage of correctly classified samples under the local roughness and local contrast feature vectors {F2, F2} using the RGB color model (block size of the
roughness are fixed on b=9)
d

R-Channel

G-Channel

B-Channel

RGB-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

83.9286

66.1232

80.9524

61.5942

83.9286

65.0362

98.8095

83.6957

82.7381

61.413

98.8095

88.587

81.5476

64.4928

83.3333

63.7681

77.9762

59.058

99.4048

89.1304

79.7619

61.0507

98.8095

86.7754

77.381

59.9638

76.7857

60.8696

75

56.8841

98.8095

83.6957

78.5714

56.7029

98.2143

83.333

73.8095

57.4275

73.2143

56.7029

73.8095

51.087

95.2381

80.0725

74.1048

51.087

95.2381

80.6159

78.5714

53.6232

75.5952

52.1739

67.8571

47.2826

94.0476

74.2754

77.381

53.9855

96.4286

75.3623

Table 8: The percentage of correctly classified samples under the local roughness and local contrast feature vectors {F2, F2} using the Lab color model (block size of the
roughness are fixed on b=9)
d

L-Channel

a-Channel

b-Channel

Lab-Channels

gray-Channel

All-Channels

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

Training

Testing

84.5238

68.8406

84.5238

69.7464

88.0952

77.7174

100

92.2101

82.7381

61.413

97.619

88.5643

83.3333

65.0362

84.5238

69.7464

89.2857

78.9855

99.4048

90.0362

79.7619

61.0507

98.2143

90.942

74.4048

59.7826

79.1667

68.6594

84.5238

72.1014

97.619

88.4058

78.5714

56.7029

98.2143

89.6739

77.9762

58.6957

83.3333

61.5942

86.3095

74.6377

97.619

85.1449

74.1048

51.087

98.2143

89.3116

75.5952

55.7971

80.9524

63.587

83.9286

72.1014

95.8333

83.5145

77.381

53.9855

98.2143

83.5145

Table 9: The percentage of correctly classified samples under the local roughness and local contrast feature vectors {F2, F2} using the HSV color model (block size of the
roughness are fixed on b=9)
d

H-Channel

S-Channel
Training

V-Channel

Testing

Training

Testing

HSV-Channels
Testing

Training

All-Channels

Training

Testing

82.7381

58.8768

80.9524

63.0435

71.4286

54.8913

97.0238

82.2464

82.7381

61.413

96.4286

85.6884

76.1905

61.413

76.1905

58.8768

67.2619

55.0725

94.6429

77.3551

79.7619

61.0507

95.8333

86.0507

72.619

57.6087

72.619

54.1667

67.2619

52.3551

89.2857

69.7464

78.5714

56.7029

97.619

78.442

81.5476

48.0072

64.881

48.7319

62.5

47.4638

93.4524

61.2319

74.4048

51.087

94.0476

71.9203

75.5952

52.8986

70.2381

46.0145

65.4762

47.2826

82.1429

57.6087

77.381

53.9855

95.8333

64.8551

Table 10: The percentage of correctly classified samples under the local
roughness and local contrast feature vectors {F2, F2} using the RGB and Lab color
model (block size of the roughness are fixed on b=9)

Training

gray-Channel

HSV

94.7464

100

95.2899

100

Lab

Training

Testing

92.9348

99.4048

91.4855

92.2101

98.8095

90.2174

100

91.8478

100

85.8696

100

8733188

98.8095

79.7101

Training

Testing

100

Table 12: The percentage of correctly classified samples under the local
roughness and local contrast feature vectors {F2, F2} using the Lab and HSV color
model (block size of the roughness are fixed on b=9)

RGB

HSV

95.471

100

95.1087

99.4048

94.3841

99.4048

98.8095

HSV

Training

Testing

100

95.8333

99.4048

95.8333

100

95.1087

92.2101

100

92.9348

89.8551

100

90.7609

Training

Testing

100

ISSN :2394-2231

Testing

Table 13: The percentage of correctly classified samples under the local
roughness and local contrast feature vectors {F2, F2} using the RGB, Lab and
HSV color model (block size of the roughness are fixed on b=9)
Lab

Lab

Training

Table 11: The percentage of correctly classified samples under the local
roughness and local contrast feature vectors {F2, F2} using the RGB and HSV
color model (block size of the roughness are fixed on b=9)
RGB

RGB

Testing

http://www.ijctjournal.org

Page 13